项目名称: 基于生成模型的迁移学习算法研究及其应用
项目编号: No.61203297
项目类型: 青年科学基金项目
立项/批准年度: 2013
项目学科: 自动化学科
项目作者: 庄福振
作者单位: 中国科学院计算技术研究所
项目金额: 24万元
中文摘要: 本项目从生成模型的角度,对迁移学习算法进行研究。通过分析源领域与目标领域之间的共性,研究基于生成模型的深度挖掘源领域与目标领域中共享主题的迁移学习方法;针对不同学习任务之间的分类知识共享与相互促进,研究基于生成模型的分类与聚类任务统一学习的迁移学习框架。在探索迁移学习算法的工作机理方面,通过引入狄利克雷模型对数据分布的描述,研究基于狄利克雷模型的源领域与目标领域数据分布不一致性度量,并分析该度量与迁移学习算法性能之间的关系。研究基于MapReduce的并行迁移学习算法用于处理海量社交网络数据中的社区划分和链接预测等方面,并研究分布式环境下只传递中间统计变量的迁移学习算法进行隐私保护。预期在SCI 或EI 收录的国际期刊及重要学术会议上发表论文15 篇。
中文关键词: 生成模型;狄利克雷过程;矩阵分解;并行算法;社交网络
英文摘要: In this project, we first study the transfer learning algorithms from the perspective of generative models. Through the analysis of commonality between the source and target domains, we study the transfer learning algorithm sharing the common topics based on generative model. And for the Scenario different tasks can benefit from each other when learning simultaneously, we study the uniform transfer learning framework for classification and clustering tasks. Second, to analyze how the transfer learning algorithms work, we introduce the Dirichlet model to study the measures of distribution differences between the source and target domains, and investigate the relationship between distribution measures and algorithmic performance. Third, we study the MapReduce based transfer learning algorithms to handle large-scale data in social network, e.g., social community partition and link prediction, and propose the distributed transfer learning algorithms to alleviate privacy-concerning when the data are distributed. We expect to publish fifteen papers in the international journals or important academic conferences, which are indexed by SCI or EI.
英文关键词: Generative Model;Dirichlet Process;Matrix Factorization;Parallel Algorithm;Social Network